Learning Personalized and Human-Aware Robot Navigation

Abstract

Robots are increasingly moving from industrial applications into everyday human environments such as healthcare, households, and public spaces. In these interactive and personal contexts, successful human-robot interaction (HRI) critically depends on robots' abilities to interpret, reflect, and adapt to individual human preferences. Yet traditional robot navigation methods, though reliable in structured environments, generally fail to capture and reflect nuanced user preferences, resulting in suboptimal user experience, reduced trust, and limited acceptance. <br/> To address these shortcomings, this thesis presents a comprehensive approach toward personalized, learning-based robot navigation. It specifically focuses on four critical aspects: (1) efficient and intuitive collection of human preferences, (2) balancing user preference reflection with robot navigation goals, (3) deriving expressive sensor representations suitable for dynamic environments, (3) deriving expressive sensor representations suitable for dynamic environments, and (4) ensuring adaptability and transparency in HRI once deployed on a robot. <br/> First, user preferences are captured using intuitive interfaces and efficient learning frameworks. We introduce a virtual reality (VR) demonstration interface, enabling users to sketch robot navigation trajectories with high intuitiveness. The VR interface is complemented with a hybrid reinforcement learning (RL) and behavioral cloning (BC) framework that requires only few demonstrations. We confirm through a user study that the personalized controller outperforms non-aligned baseline approaches, with users reporting that their preferences were better reflected. Besides demonstration-based approaches, we also optimize the preference collection through RL from human feedback (RLHF). We introduce the novel query generation approach "EnQuery" based on policy ensembles, maximizing the information gain in low-query regimes, while providing trajectory options with common start and goal reference points. EnQuery subsequently drives a user study that compares immersive VR and conventional 2D video interfaces for preference collection. Here, we find effects of the interface modality on user experience, preference consistency, and policy alignment. <br/> Second, the thesis develops and validates learning architectures that balance the trade-off between user preference reflection and robot task completion. The proposed hybrid RL+BC learning framework internalizes user preferences while preserving goal-directed performance. To quantify the quality of preference reflection in navigation trajectories, we introduce a new metric that is based on the Fréchet Distance. <br/> Third, we address the challenge of sensor representation for robust navigation in dynamic, human-populated environments. A depth vision-based perception pipeline employing a variational autoencoder and motion prediction compresses sensor observations into latent states, capturing both scene details and the user for effective personalized policy learning. In parallel, a spatiotemporal attention mechanism paired with a novel 2D lidar state representation improves obstacle avoidance and foresight in dynamic human environments over state-of-the-art baselines. <br/> Fourth, the thesis advances the adaptability and transparency of learning-based robot navigation policies. To accommodate adaptability to evolving user preferences, a multi-objective RL framework facilitates principled post-deployment tuning of demonstration reflection and other navigation objectives. For improved transparency between robot and user, an explainable artificial intelligence (AI) interface in VR is developed, visually grounding navigation policy attribution scores semantically in scene context. The approach communicates internal decision-making of black-box neural network policies in an intuitive manner and thereby improves non-expert users' objective and subjective understanding of robot behavior. <br/> These contributions are validated through extensive simulation studies, user experiments, and real-robot deployments. The findings demonstrate that preference-reflecting, learning-based navigation is achievable, robust, and perceived as superior to classical approaches by users. The insights regarding interface modality, interaction sample efficiency, sensor abstraction, and explainability inform the design of future user-centric robotic systems. In summary, this thesis establishes principled methods for navigation preference collection, learning, and behavior explanation, advancing the state-of-the-art towards seamless, preference-aware HRI in daily life.